TaxoAttention: fusing semantic and hierarchical taxonomic models with attention-guided augmentation for mosquito identification
Accurate species-level identification of mosquitoes remains a critical bottleneck in vector surveillance, particularly when dealing with morphologically similar taxa or field-damaged specimens. In this study, we introduce a deep learning framework that integrates hierarchical taxonomic structure with attention-based feature learning to improve classification performance under real-world imaging conditions. The model combines a semi-fixed, taxonomy-informed feature stream with a fully trainable semantic stream. It further employs dynamic attention-guided augmentation, designed to focus training on diagnostically relevant regions, such as wing venation, tarsal banding, and palpal morphology, and to reduce the influence of the background. Evaluated on a standardized, predominantly field-collected dataset (MosquitoTax-200) representing 32 species across Aedes, Anopheles, and Culex with 200 images per species, our approach achieves 84.6% macro-averaged precision (mAP), 5.0 percentage points above an Xception baseline trained with standard cross-entropy loss and used as a control. This work highlights the potential of biologically grounded AI tools to scale species identification in vector control operations and to support training and quality assurance in entomological diagnostics.
Authors
- Roy Faiman (ORCID: https://orcid.org/0000-0002-9997-7801)
- Sameerah Talafha (ORCID: https://orcid.org/0000-0002-5302-8539)
- Atharva Mhaskar
Institutions
- Vectec Solutions (United States) (US)
- Carnegie Mellon University (US)
Publication Details
- Journal
- Frontiers in Malaria
- Published
- 2026-11-13
- DOI
- https://doi.org/10.3389/fmala.2026.1978636
- Primary Topic
- Advanced Neural Network Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- National Science Foundation